You are on page 1of 5

Nonlinear Biomedical Physics BioMed Central

Research Open Access

Everything you wanted to ask about EEG but were afraid to get the
right answer
Wlodzimierz Klonowski

Address: Institute of Biocybernetics and Biomedical Engineering, Polish Academy of Sciences, Warsaw, Poland
Email: Wlodzimierz Klonowski -

Published: 26 May 2009 Received: 17 February 2009

Accepted: 26 May 2009
Nonlinear Biomedical Physics 2009, 3:2 doi:10.1186/1753-4631-3-2
This article is available from:
© 2009 Klonowski; licensee BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

We answer several important questions concerning EEG. We also shortly discuss importance of
nonlinear methods of contemporary physics in EEG analysis. Basic definitions and explanation of
fundamental concepts may be found in my previous publications in NBP.
It is a magnificent feeling to recognize the unity of complex phenomena which appear to be things quite
apart from the direct visible truth.
Albert Einstein

Q1. Why anybody might be afraid of getting the nals are resultant of activity of many brain cells – the exist-
right answers about EEG? ence of a particular frequency in the FFT does not
A1. Habit is a second nature. Before personal computers necessarily mean that there is an oscillator in the brain at
came into in medicine in 1980's EEG signals were regis- that frequency. New nonlinear methods should be much
tered on a wide paper tape When EEG was registered on a more appropriate for EEG-analysis but are still used only
paper tape vertical lines 3 cm away, moving 1.5, 3.0, or 6.0 in research and not in everyday clinical practice because
cm/sec depending on the system, doctor who interpreted they are not yet rooted in medical tradition and they
this EEG could easily observe frequency of EEG waves by meet strong barrier of doctors' habits. Somebody who
counting number of pen sways (wave ridges) in one sec- got accustomed to spectral methods does not want to
ond if there were, say, from 3 up to 30 wave maxima learn new methods not even to ask questions the right
between two vertical lines. From this paper-tape period answers to which could make the person to understand
come classical names of EEG-bands, in particular α and β and admit that the methods he/she has been using might
bands; if there were waves of small frequency they could be often wrong or at least misleading. Basic definitions
be too flat to be observed by naked eye, while those with and explanation of fundamental concepts may be found
frequencies higher than 30 merged together due to the in many papers and books [1,2].
width of pen's line and were considered to be just a noise.
So, when development of personal computers enabled Q2. Is brain deterministic or stochastic?
numerical registration of EEG-signals and their spectral A2. Human brain may neither act as a fully deterministic
analysis using linear methods like FFT developed at the system because it would not be able to invent anything
same time, Medical Doctors accepted those methods quite new, nor may it act as a fully stochastic system because it
easily and even the new 'slow' and 'quick' brainwaves would not be able to learn and exactly repeat any
bands were introduced. It does not matter that EEG-sig- sequence of thoughts. The answer to this dilemma lies in

Page 1 of 5
(page number not for citation purposes)
Nonlinear Biomedical Physics 2009, 3:2

nonlinear dynamics and deterministic chaos – human but rather it is a reflection of switching of the inherent
brain is a complex nonlinear system showing compli- metastable states of neural assemblies during brain func-
cated emergent properties, including consciousness. It tioning. EEG-signal recorded from a scalp electrode is
is much more appropriate to use nonlinear methods for influenced by different deeper brain structures, each
analysis of signals generated by such a complex nonlinear 'transmitting' with different and changeable intensity; so,
system, despite the fact that for short time intervals linear in a fraction of a second the main source of the registered
methods like FFT may work o.k. Nonlinear methods may signal often moves from one brain structures to another.
be applied to linear signals – one might try to approxi- And if source of a signal changes with time then the signal
mate straight line using parabolic function only to find is obviously nonstationary. Nonstationarity arises also
that coefficient of the quadratic term was practically equal because of different time scales involved in the dynamical
zero. It is the opposite that obviously fails – if one uses lin- process – dynamical parameters are sensitive to the time
ear approximation then one will never be able to appro- scales and hence in the study of brain one must identify
priately approximate a nonlinear function on a all relevant time scales involved in the process to get an
sufficiently long interval, but on a short interval linear insight in the working of brain [5]. It is extremely impor-
predictability may work o.k. It is unbelievable but in the tant that fractal methods easily detect nonstationarities in
XXI century some scientists still maintain they have dem- the analyzed signals, nonstationarities that are not easily
onstrated by using methods like linear forecasting or sur- detectable by linear methods like FFT [1]. Nonstationari-
rogate data tests of EEG time series that EEG is linear ties in EEG are also due to pathological changes, for exam-
signal as if the human brain is a linear system. Although ple epileptic seizures, or to changes of the physiological
linear systems may have a small range of applicability, it state, for example passing from one sleep stage to another.
is inappropriate to use a linear system to deal with the
highly nonlinear complexity of the brain. Perfect linear Q5. Is it necessary to apply surrogate data
system is a kind of abstraction like an ideal gas. method to EEG time series?
A5. Surrogate data tests may demonstrate that it is not
Q3. Is EEG a linear signal? impossible that the analyzed signal is generated by a lin-
A3. No, EEG-signal generated by brain is nonlinear. ear Gaussian process, but such tests may never prove that
While practically all researchers do agree that human this is really the case. On the other hand, it is obvious that
brain is the most complex system we know, many signals generated by human brain, the most complex sys-
researchers still claim that brain is linear, or at least that tem we know, are really complex nonlinear signals and
EEG-signals generated by brain are linear, without realiz- that there exist much more general premises to assume
ing that these 'believes' contradict one another. Many bio- that EEG-signals are nonlinear (cf. A2. above) and linear
medical researchers are 'infected with Human Linearity methods may eventually be used only for sufficiently
Virus' (HLV) – they 'think linearly' and ignore the facts short time-scale. For example, if an organism is placed
that human body, and, in particular, human brain are under the influence of ionizing radiation for a short
complex nonlinear systems generating nonstationary period i.e. if only a small dose of radiation is absorbed the
nonlinear signals, and that appropriate analysis of such effect exerted on the organism may be positive, propor-
signals does need new nonlinear methods. We have dem- tional to the absorbed dose if that was sufficiently small,
onstrated an example of astonishing similarity between while for longer periods of exposition and higher
economic crisis and epileptic seizure in the brain if appro- absorbed doses the effect becomes strongly negative and
priate nonlinear methods (like Higuchi's fractal dimen- strongly nonlinear, leading to the death of the organism;
sion) are applied to the analysis of a 'signal' generated by this phenomenon is called radiation hormesis [1,6]. When
an 'economic organism' – time series of Dow Jones index analyzing systems that show hormetic effects we clearly
during the period of 'big crash' – and EEG-signal during have to look for nonlinear dose-effect relation. Similarly,
epileptic seizure [1,3]. even before we were able to have a look onto the Earth
from cosmic space to affirm that it really has spherical
Q4. Is EEG a stationary signal? shape there were enough of other premises to affirm the
A4. No, EEG signal is nonstationary. In general, biosig- assumption of Earth's spherical shape. Even if a geodesist
nals are '3N' – Nonstationary, Nonlinear, Noisy. Nonsta- had measured many triangles on the Earth surface and
tionarity means that signal's statistical characteristics found that the sum of angles in each of those triangles was
change with time. The brain activity is essentially nonsta- not statistically different from 180° (that is of course true
tionary. Quasi-stationary segments in EEG have duration for triangles which sides are short in comparison with
about 0.25 sec [4]. The basic source of the observed non- Earth's radius) nobody would have maintained that it was
stationarity in EEG signal is not due to the casual influ- the evidence that the Earth is flat. What seems to be true
ences of the external stimuli on the brain mechanisms, on comparatively short scale (in time or in space) may no

Page 2 of 5
(page number not for citation purposes)
Nonlinear Biomedical Physics 2009, 3:2

longer be true on longer scale (cf. A3. above) – a neurosci- on spectral linear analysis of biosignals like EEG. Human
entists who maintains that EEG is linear is like a geodesists organism (and human brain in particular) is a highly
who maintains that Earth is flat. complex nonlinear systems, and that is why standardized
approach based on 'stiff' protocols may lead to serious
Q6. Is the routine EEG-acquisition done errors of judgment. There are differences in defining 'nor-
properly? mal' ranges even between quite reliable sources. For exam-
A6. No, routine EEG-acquisition is done improperly ple, according to NLM 'Normally, the ICP [Intracranial
from the point of view of neuroscientific research. It is Pressure] ranges from 1 to 15 mm Hg' [8] but other
obvious that the higher is the frequency of a wave the sources give ranges like 8 to 18 mm Hg; anyway, what for
more information it may carry, but frequencies exceeding one person is a quite high ICP for another may be quite
roughly 70–90 Hz are filtered out by EEG-data acquisition low. Similarly, nothing like 'normal EEG' exists, one
systems hardware, and so are very low frequencies smaller may only compare EEG of the same person in different
than 0.5 Hz. It is so due to paper tape tradition and doc- time periods, in particular 'before' and 'after', e.g. compar-
tors' habits (cf. A1. above). On the other hand, neurosci- ing EEG before and after administration of different drugs.
entists have found that some tasks, like e.g. face Markers supplied by Biomedical Physicists, e.g. quantita-
perception, elicit frequencies up to 250 Hz when the stim- tive descriptors of EEG-signal adapted from Nonlinear
uli are processed by human brains [7]. Most of manufac- Dynamics, may help in better assessment of various spon-
turers of EEG apparata do not leave any possibility for taneous or evoked, normal and pathological functional
users to register unfiltered (raw) EEG-signal for further states of the brain in neuropsychiatric patients, and so
analysis with new analytical methods. Only recently some may be helpful in deciding diagnosis, treatment, and
manufacturers like NeuroConn GmbH (Ilmenau, Ger- prognosis.
many) supply EEG-systems that enable registration of
unfiltered signals. Q8. Is it possible to model EEG-signals with
systems of ordinary differentials equations
Q7. What is 'normal EEG'? (ODE)?
A7. In Medicine two plus two not always equals four and A8. Using systems of ODE with several parameters that
that is why there is no alternative to Personalized Medi- may not exactly be calculated from carefully done experi-
cine provided by well trained Medical Doctors. The nor- ments one may modelled practically any process. For
mal body temperature 36.6°C is the norm, but it is not example, Liley's model consists of 14 first order nonlinear
easy to point to any other norm like that. It is erroneous ODE with 29 physiological and anatomical parameters,
belief that in Medicine the normal value equals the pop- some of which may take values from continuous ranges of
ulation average value – so called 'normative databases' of possible values [9]. Using numerical computer methods
EEG, no matter how large, do not give a possibility of 'reli- one may solve such systems of ODE which might as well
able comparison' to decide if the given case is 'normal' or model 'problems' like 'influence of the baldness of the
'abnormal'. Moreover, those databases are mostly based forward basketball players on the results of NBL
matches'. One may even predict results of several consec-
utive matches based on such a model. To characterize
baldness one needs several parameters, plus one needs
more obvious but not exactly measurable parameters
characterizing momentary physical fitness of all players,
etc. There exist infinite number of combinations of ODE
parameters that will lead to the same prediction on some
short interval, but will differ tremendously on a longer
time interval. So, such models do not explain practically
anything nor they give a possibility to measure some
parameters of the system under consideration. Even a
model consisting of only 3 simple quasilinear ODE with
3 parameters (so called Lorenz equations, [1]) of which 2
with frequency
–a one
of frequency
1 Hz (upper)
12 Hz
Hz results
com- are constant while 1 slightly changes may lead to quite
Fourier decomposition of a signal of frequency 12 Hz unexpected behaviour. These remarks concern not only
with the amplitude modulated with frequency 1 Hz modelling of EEG-signals but practically any ODE model
(upper) results in two harmonic signals – one with with several parameters that should be found experimen-
frequency 11 Hz and another with frequency 13; the tally, like ECG-signals, membrane transport, sugar level
basic frequency 12 Hz completely disappears. regulation, etc, etc.

Page 3 of 5
(page number not for citation purposes)
Nonlinear Biomedical Physics 2009, 3:2

a. c.

b. d.
2 toc.twoandsimilar
d.) – stationary (upper left, a.) and nonstationary (lower left, b.) signals gives dramatically different
FFT applied to two similar – stationary (upper left, a.) and nonstationary (lower left, b.) signals gives dramati-
cally different results (right, c. and d.), while application of Higuchi's algorithm gives quite similar values of average fractal
dimension, Df, of both signals. The stationary signal was composed of five harmonic waves of different frequencies, and then
randomly chosen small segments were removed from that stationary signal, so forming a nonstationary signal that when
decomposed using Fourier leads to a very 'rich' spectrum of frequencies. Routine artefacts’ correction in EEG-signals may lead
to similar unreliable cases.

Q9. What are shortcomings of linear methods? ing a sum of two sine functions into a product of sine and
A9. Linear methods like FFT (Fast Fourier Transform), WT cosine that could be easily computed with a slide (loga-
(wavelet transform), or MP (Matching Pursuit) work rithmic) rule (Figure 1). We do need new nonlinear meth-
properly only for stationary signals but assumptions of ods of biosignal analysis; otherwise while living in XXI
stationarity required for the correct use of these algo- century, we will still be plunged in XIX century 'linear sci-
rithms are often ignored. WT and MP has better accuracy ence' of Fourier and Markov.
than FFT but much bigger ambiguity in signal decomposi-
tion. Linear methods may lead to very misleading results. Q10. What are advantages of nonlinear
E.g. if in a measured signal one observes regular waves of methods?
frequency 12 Hz with amplitude modulated with fre- A10. One may apply methods of nonlinear analysis, for
quency 1 Hz, then Fourier decomposition of this signal example Higuchi's fractal dimension method, Df [1,10],
leads to two components, each of amplitude equal half of to any signal. Df's value, always between 1.0 and 2.0, is
that of the analyzed signal, with frequencies 11 Hz and 13 just a measure of what is called signal's complexity. Unlike
Hz respectively, while the basic frequency of the analyzed fractal dimension in phase space [1]Df is calculated
signal (12 Hz) does not appear at all in the Fourier spec- directly in time domain; running fractal dimension, Df (t),
trum; such a result may be predicted even without any cal- may be calculated using a moving window as short as 70–
culations, from a simple trigonometric formula that was 100 data points. It is not necessary to made surrogate data
used in high schools in 'pre-computer era' for transform- test before applying Higuchi's fractal dimension method

Page 4 of 5
(page number not for citation purposes)
Nonlinear Biomedical Physics 2009, 3:2

because it does not matter if the analyzed signal itself is

'really chaotic' – it may be deterministic, stochastic, non-
stationary and noisy. Moreover, generation of surrogate
data often applies linear transformations like FFT and
reverse FFT that may show serious shortcomings. E.g. FFT
applied to similar stationary and nonstationary signals
gives dramatically different results, while application of
Higuchi's algorithm gives quite similar values of average
fractal dimension, Df, of both signals (Figure 2). Df is lin-
early related to so called Hurst exponent that is a measure
of the tendency for time series values to persist or to alter-
nate and works on prediction of chaotic time series sug-
gest that humans are sensitive to the Hurst exponent [11],
and so must be sensitive also to Df.

We all become more and more specialized in very narrow
disciplines and we often do not know that the methods
we want to apply in our research have been used for a long
time in other disciplines. When we learn about it we are
often amazed like Molier's Mr. Jourdain (Le Bourgeois Gen-
tilhomme II. iv) who says: 'Good heaven! For more than
forty years I have been speaking prose without knowing
it'. Our philosophy is that to be applicable a method
should preferably be really simple and easily understand-
able to non-specialists in the field. Some nonlinear meth-
ods like Higuchi's fractal dimension method are very
simple – they draw from multiple disciplines and have
multidisciplinary applications.

Our work was supported by IBBE PAS by 4.4/st/09.

1. Klonowski W: From conformons to human brains: an informal
overview of nonlinear dynamics and its applications in bio-
medicine. Nonlinear Biomedical Physics 2007, 1:5.
2. Sanei S, Chambers JA: EEG signal processing Wiley, New York, NY,
USA; 2007.
3. Klonowski W, Olejarczyk E, Stepien R: 'Epileptic seizures' in eco-
nomic organism. Physica A 2004, 342:701-707.
4. Kaplan AY, Fingelkurts AA, Fingelkurts AA, Borisov SV, Darkhovsky
BS: Neuronal coordination in the brain: A signal processing
perspective. Signal processing 2005, 85:2190-2212.
5. Indic P, Pratap R, Nampoori VP, Pradhan N: Significance of time
scales in nonlinear dynamical analysis of electroencephalo-
gram signals. Int J Neurosci 1999, 99(1-4):181-194.
6. Jaworowski Z: Beneficial Effects of Radiation and Regulatory
Policy. Australas Phys Eng Sci Med 1997, 20(3):125-138. Publish with Bio Med Central and every
7. Lachaux JP, et al.: Workshop on 'Oscillations and higher brain function', scientist can read your work free of charge
Lyon 2003 [].
"BioMed Central will be the most significant development for
8. Medline Plus, Intracranial pressure monitoring: [http://]. disseminating the results of biomedical researc h in our lifetime."
9. Bojak I, Liley DT: Modeling the effects of anesthesia on the Sir Paul Nurse, Cancer Research UK
electroencephalogram. Phys Rev E Stat Nonlin Soft Matter Phys
2005, 71(4 Pt 1):041902. Your research papers will be:
10. Higuchi T: Approach to an irregular time series on the basis of available free of charge to the entire biomedical community
the fractal theory. Physica D 1988, 31:277-283.
11. Smithson M: Judgment under chao. Organizational Behavior and peer reviewed and published immediately upon acceptance
Human Decision Processes 1997, 69:59-66. cited in PubMed and archived on PubMed Central
yours — you keep the copyright

Submit your manuscript here: BioMedcentral

Page 5 of 5
(page number not for citation purposes)